Track memory allocation with restrack DB help (Part II)
Subsystems (3)
v5 cover Track memory allocation with restrack DB help (Part II)
From: Leon Romanovsky <[email protected]>
Changelog:
v5:
* Reorder patches to postpone changes in rdma_restrack_add to be in next series.
v4: https://lore.kernel.org/linux-rdma/[email protected]/
* Rebased on latest for-upstream, all that time the patches were in
our regression and didn't introduce any issues.
* Took first five patches that hadn't any comments
v3: https://lore.kernel.org/lkml/[email protected]
* Rebased on already accepted patches.
* Added mlx4 special QPs to the list of not-tracked QPs (dropped previous mlx4 special QP patch).
* Separated to two patches change in return value of cma_listen_* routines.
* Changed commit messages and added Fixes as Jason requested.
v2: https://lore.kernel.org/linux-rdma/[email protected]/
* Added new patch to fix mlx4 failure on SR-IOV, it didn't have port set.
* Changed "RDMA/cma: Delete from restrack DB after successful destroy" patch.
v1: https://lore.kernel.org/lkml/[email protected]
* Fixed rebase error, deleted second assignment of qp_type.
* Rebased code on latests rdma-next, the changes in cma.c caused to change
in patch "RDMA/cma: Delete from restrack DB after successful destroy".
* Dropped patch of port assignment, it is already done as part of this
series.
* I didn't add @calller description, regular users should not use _named() funciton.
v0: https://lore.kernel.org/lkml/[email protected]
----------------------------------------------------------------------------------
Leon Romanovsky (3):
RDMA/core: Track device memory MRs
RDMA/core: Allow drivers to disable restrack DB
RDMA/restrack: Support all QP types
drivers/infiniband/core/core_priv.h | 25 ++++++------------- drivers/infiniband/core/counters.c | 2 +- drivers/infiniband/core/restrack.c | 12 +++++++-- drivers/infiniband/core/uverbs_cmd.c | 4 +-- drivers/infiniband/core/uverbs_std_types_mr.c | 4 +++ drivers/infiniband/core/uverbs_std_types_qp.c | 4 +-- drivers/infiniband/core/verbs.c | 11 ++++---- drivers/infiniband/hw/mlx4/qp.c | 5 ++++ drivers/infiniband/hw/mlx5/qp.c | 2 +- include/rdma/ib_verbs.h | 10 ++++++-- include/rdma/restrack.h | 24 ++++++++++++++++++ 11 files changed, 70 insertions(+), 33 deletions(-)
--
2.28.0
PATCH v5 1/3 RDMA/core: Track device memory MRs
From: Leon Romanovsky <[email protected]>
Device memory (DM) are registered as MR during initialization flow, these MRs were not tracked by resource tracker and had res->valid set as a false. Update the code to manage them too.
Before this change:
[leonro@vm ~]$ ibv_rc_pingpong -j &
[leonro@vm ~]$ rdma res show mr <-- shows nothing
After this change:
[leonro@vm ~]$ ibv_rc_pingpong -j &
[leonro@vm ~]$ rdma res show mr
dev ibp0s9 mrn 0 mrlen 4096 pdn 3 pid 734 comm ibv_rc_pingpong
--- drivers/infiniband/core/uverbs_std_types_mr.c | 4 ++++ 1 file changed, 4 insertions(+)
drivers/infiniband/core/uverbs_std_types_mr.c +4 -0
@@ -33,6 +33,7 @@#include "rdma_core.h" #include "uverbs.h" #include <rdma/uverbs_std_types.h> +#include "restrack.h" static int uverbs_free_mr(struct ib_uobject *uobject, enum rdma_remove_reason why,
@@ -134,6 +135,9 @@ static int UVERBS_HANDLER(UVERBS_METHOD_DM_MR_REG)(atomic_inc (&pd->usecnt);atomic_inc (&dm->usecnt); +rdma_restrack_new (&mr->res,RDMA_RESTRACK_MR ); +rdma_restrack_set_name (&mr->res,NULL ); +rdma_restrack_add (&mr->res); uobj->object = mr;uverbs_finalize_uobj_create (attrs,UVERBS_ATTR_REG_DM_MR_HANDLE );
Signature · 1 line
2.28.0
PATCH v5 2/3 RDMA/core: Allow drivers to disable restrack DB
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
Driver QP types are special case with no IBTA restrictions. For example, EFA implemented creation of this QP type as regular one, while mlx5 separated create to two step: create and modify. That separation causes to the situation where DC QP (mlx5) is always added to the same xarray index zero.
This change allows to drivers like mlx5 simply disable restrack DB tracking, but it doesn't disable kref on the memory.
--- drivers/infiniband/core/counters.c | 2 +- drivers/infiniband/core/restrack.c | 12 ++++++++++-- drivers/infiniband/hw/mlx4/qp.c | 5 +++++ drivers/infiniband/hw/mlx5/qp.c | 2 +- include/rdma/restrack.h | 24 ++++++++++++++++++++++++ 5 files changed, 41 insertions(+), 4 deletions(-)
drivers/infiniband/core/counters.c +1 -1
@@ -285,7 +285,7 @@ int rdma_counter_bind_qp_auto(struct ib_qp *qp, u8 port)struct rdma_counter *counter;int ret; -if (!qp->res.valid ||rdma_is_kernel_res (&qp->res)) +if (!rdma_restrack_is_tracked (&qp->res) ||rdma_is_kernel_res (&qp->res))return 0 ;if (!rdma_is_port_valid (dev, port))
drivers/infiniband/core/restrack.c +10 -2
@@ -221,11 +221,14 @@ void rdma_restrack_add(struct rdma_restrack_entry *res) {struct ib_device *dev =res_to_dev (res);struct rdma_restrack_root *rt; -int ret; +int ret =0 ;if (!dev)return ; +if (res->no_track) + goto out; + rt = &dev->res[res->type];if (res->type ==RDMA_RESTRACK_QP ) {
@@ -253,6 +256,7 @@ void rdma_restrack_add(struct rdma_restrack_entry *res) &rt->next_id, GFP_KERNEL); } +out:if (!ret) res->valid = true; }
@@ -325,6 +329,9 @@ void rdma_restrack_del(struct rdma_restrack_entry *res)return ; } +if (res->no_track) + goto out; + dev =res_to_dev (res);if (WARN_ON (!dev))return ;
@@ -335,8 +342,9 @@ void rdma_restrack_del(struct rdma_restrack_entry *res)if (res->type ==RDMA_RESTRACK_MR || res->type ==RDMA_RESTRACK_QP )return ;WARN_ON (old != res); - res->valid = false; +out: + res->valid = false;rdma_restrack_put (res);wait_for_completion (&res->comp); }
drivers/infiniband/hw/mlx4/qp.c +5 -0
@@ -1561,6 +1561,11 @@ static int _mlx4_ib_create_qp(struct ib_pd *pd, struct mlx4_ib_qp *qp,if (err)return err; +if (init_attr->create_flags & + (MLX4_IB_SRIOV_SQP |MLX4_IB_SRIOV_TUNNEL_QP )) +/* Internal QP created with ib_create_qp */ +rdma_restrack_no_track (&qp->ibqp.res); + qp->port = init_attr->port_num; qp->ibqp.qp_num = init_attr->qp_type ==IB_QPT_SMI ?0 : init_attr->create_flags &MLX4_IB_QP_CREATE_ROCE_V2_GSI ? sqpn :1 ;
drivers/infiniband/hw/mlx5/qp.c +1 -1
@@ -2410,7 +2410,7 @@ static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd, } qp->state =IB_QPS_RESET ; - +rdma_restrack_no_track (&qp->ibqp.res);return 0 ; }
include/rdma/restrack.h +24 -0
@@ -68,6 +68,14 @@ struct rdma_restrack_entry {* As an example for that, see mlx5 QPs with type MLX5_IB_QPT_HW_GSI */ bool valid; +/** +* @no_track: don't add this entry to restrack DB +* +* This field is used to mark an entry that doesn't need to be added to +* internal restrack DB and presented later to the users at the nldev +* query stage. +*/ + u8 no_track : 1;/* * @kref: Protect destroy of the resource */
@@ -145,4 +153,20 @@ int rdma_nl_stat_hwcounter_entry(struct sk_buff *msg, const char *name,struct rdma_restrack_entry *rdma_restrack_get_byid (struct ib_device *dev,enum rdma_restrack_type type,u32 id); + +/** +* rdma_restrack_no_track() - don't add resource to the DB +* @res: resource entry +* +* Every user of thie API should be cross examined. +* Probaby you don't need to use this function. +*/ +static inline void rdma_restrack_no_track (struct rdma_restrack_entry *res) +{ + res->no_track = true; +} +static inline bool rdma_restrack_is_tracked (struct rdma_restrack_entry *res) +{ +return !res->no_track; +}#endif /* _RDMA_RESTRACK_H_ */
Signature · 1 line
2.28.0
PATCH v5 3/3 RDMA/restrack: Support all QP types
In reply to Leon RomanovskyFrom: Leon Romanovsky <[email protected]>
The latest changes in restrack name handling allowed to simplify the QP creation code to support all types of QPs.
For example XRC QP are presented with inbox rdmatool.
[leonro@vm ~]$ ibv_xsrq_pingpong &
[leonro@vm ~]$ rdma res show qp
link ibp0s9/1 lqpn 0 type SMI state RTS sq-psn 0 comm [ib_core]
link ibp0s9/1 lqpn 1 type GSI state RTS sq-psn 0 comm [ib_core]
link ibp0s9/1 lqpn 7 type UD state RTS sq-psn 0 comm [mlx5_ib]
link ibp0s9/1 lqpn 42 type XRC_TGT state INIT sq-psn 0 path-mig-state MIGRATED comm [ib_uverbs] link ibp0s9/1 lqpn 43 type XRC_INI state INIT sq-psn 0 path-mig-state MIGRATED pdn 197 pid 419 comm ibv_xsrq_pingpong
--- drivers/infiniband/core/core_priv.h | 25 ++++++------------- drivers/infiniband/core/uverbs_cmd.c | 4 +-- drivers/infiniband/core/uverbs_std_types_qp.c | 4 +-- drivers/infiniband/core/verbs.c | 11 ++++---- include/rdma/ib_verbs.h | 10 ++++++-- 5 files changed, 25 insertions(+), 29 deletions(-)
drivers/infiniband/core/core_priv.h +7 -18
@@ -318,15 +318,12 @@ struct ib_device *ib_device_get_by_index(const struct net *net, u32 index);void nldev_init (void );void nldev_exit (void ); -static inline struct ib_qp *_ib_create_qp (struct ib_device *dev, -struct ib_pd *pd, -struct ib_qp_init_attr *attr, -struct ib_udata *udata, -struct ib_uqp_object *uobj) +static inline struct ib_qp * +_ib_create_qp (struct ib_device *dev,struct ib_pd *pd, +struct ib_qp_init_attr *attr,struct ib_udata *udata, +struct ib_uqp_object *uobj,const char *caller) { -enum ib_qp_type qp_type = attr->qp_type;struct ib_qp *qp; -bool is_xrc;if (!dev->ops.create_qp)return ERR_PTR (-EOPNOTSUPP );
@@ -347,7 +344,6 @@ static inline struct ib_qp *_ib_create_qp(struct ib_device *dev, qp->srq = attr->srq; qp->rwq_ind_tbl = attr->rwq_ind_tbl; qp->event_handler = attr->event_handler; - qp->qp_type = attr->qp_type; qp->port = attr->port_num;atomic_set (&qp->usecnt,0 );
@@ -356,16 +352,9 @@ static inline struct ib_qp *_ib_create_qp(struct ib_device *dev,INIT_LIST_HEAD (&qp->sig_mrs);rdma_restrack_new (&qp->res,RDMA_RESTRACK_QP ); -/* -* We don't track XRC QPs for now, because they don't have PD -* and more importantly they are created internaly by driver, -* see mlx5 create_dev_resources() as an example. -*/ - is_xrc = qp_type ==IB_QPT_XRC_INI || qp_type ==IB_QPT_XRC_TGT ; -if ((qp_type <IB_QPT_MAX && !is_xrc) || qp_type ==IB_QPT_DRIVER ) { -rdma_restrack_parent_name (&qp->res, &pd->res); -rdma_restrack_add (&qp->res); - } +WARN_ONCE (!udata && !caller,"Missing kernel QP owner" ); +rdma_restrack_set_name (&qp->res, udata ?NULL : caller); +rdma_restrack_add (&qp->res);return qp; }
drivers/infiniband/core/uverbs_cmd.c +2 -2
@@ -1400,8 +1400,8 @@ static int create_qp(struct uverbs_attr_bundle *attrs,if (cmd->qp_type ==IB_QPT_XRC_TGT ) qp =ib_create_qp (pd, &attr);else - qp =_ib_create_qp (device, pd, &attr, &attrs->driver_udata, - obj); + qp =_ib_create_qp (device, pd, &attr, &attrs->driver_udata, obj, +NULL );if (IS_ERR (qp)) { ret =PTR_ERR (qp);
drivers/infiniband/core/uverbs_std_types_qp.c +2 -2
@@ -251,8 +251,8 @@ static int UVERBS_HANDLER(UVERBS_METHOD_QP_CREATE)(if (attr.qp_type ==IB_QPT_XRC_TGT ) qp =ib_create_qp (pd, &attr);else - qp =_ib_create_qp (device, pd, &attr, &attrs->driver_udata, - obj); + qp =_ib_create_qp (device, pd, &attr, &attrs->driver_udata, obj, +NULL );if (IS_ERR (qp)) { ret =PTR_ERR (qp);
drivers/infiniband/core/verbs.c +6 -5
@@ -1191,7 +1191,7 @@ static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp, }/** -* ib_create_qp - Creates a kernel QP associated with the specified protection +* ib_create_named_qp - Creates a kernel QP associated with the specified protection * domain. * @pd: The protection domain associated with the QP. * @qp_init_attr: A list of initial attributes required to create the
@@ -1200,8 +1200,9 @@ static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,* * NOTE: for user qp use ib_create_qp_user with valid udata! */ -struct ib_qp *ib_create_qp (struct ib_pd *pd, -struct ib_qp_init_attr *qp_init_attr) +struct ib_qp *ib_create_named_qp (struct ib_pd *pd, +struct ib_qp_init_attr *qp_init_attr, +const char *caller) {struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;struct ib_qp *qp;
@@ -1226,7 +1227,7 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,if (qp_init_attr->cap.max_rdma_ctxs)rdma_rw_init_qp (device, qp_init_attr); - qp =_ib_create_qp (device, pd, qp_init_attr,NULL ,NULL ); + qp =_ib_create_qp (device, pd, qp_init_attr,NULL ,NULL , caller);if (IS_ERR (qp))return qp;
@@ -1292,7 +1293,7 @@ struct ib_qp *ib_create_qp(struct ib_pd *pd,return ERR_PTR (ret); } -EXPORT_SYMBOL (ib_create_qp); +EXPORT_SYMBOL (ib_create_named_qp); static const struct { int valid;
include/rdma/ib_verbs.h +8 -2
@@ -3618,8 +3618,14 @@ static inline int ib_post_srq_recv(struct ib_srq *srq, bad_recv_wr ? : &dummy); } -struct ib_qp *ib_create_qp (struct ib_pd *pd, -struct ib_qp_init_attr *qp_init_attr); +struct ib_qp *ib_create_named_qp (struct ib_pd *pd, +struct ib_qp_init_attr *qp_init_attr, +const char *caller); +static inline struct ib_qp *ib_create_qp (struct ib_pd *pd, +struct ib_qp_init_attr *init_attr) +{ +return ib_create_named_qp (pd, init_attr,KBUILD_MODNAME ); +}/** * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
Signature · 1 line
2.28.0
Quoted 26 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Changelog:
v5:
* Reorder patches to postpone changes inrdma_restrack_addto be in next series.
v4: https://lore.kernel.org/linux-rdma/[email protected]/
* Rebased on latest for-upstream, all that time the patches were in
our regression and didn't introduce any issues.
* Took first five patches that hadn't any comments
v3: https://lore.kernel.org/lkml/[email protected]
* Rebased on already accepted patches.
* Added mlx4 special QPs to the list of not-tracked QPs (dropped previous mlx4 special QP patch).
* Separated to two patches change in return value of cma_listen_* routines.
* Changed commit messages and added Fixes as Jason requested.
v2: https://lore.kernel.org/linux-rdma/[email protected]/
* Added new patch to fix mlx4 failure on SR-IOV, it didn't have port set.
* Changed "RDMA/cma: Delete from restrack DB after successful destroy" patch.
v1: https://lore.kernel.org/lkml/[email protected]
* Fixed rebase error, deleted second assignment ofqp_type.
* Rebased code on latests rdma-next, the changes incma.ccaused to change
in patch "RDMA/cma: Delete from restrack DB after successful destroy".
* Dropped patch of port assignment, it is already done as part of this
series.
* I didn't add @calller description, regular users should not use_named()funciton.
v0: https://lore.kernel.org/lkml/[email protected]
Quoted 6 lines
----------------------------------------------------------------------------------
Leon Romanovsky (3):
RDMA/core: Track device memory MRs
RDMA/core: Allow drivers to disable restrack DB
RDMA/restrack: Support all QP types
Applied to for-next, thanks
Jason